A New YOLO Based Spatio-Behavioural Feature Analysis for Anomaly Detection in Crowd Videos

Abstract

This study presents a hybrid deep learning model that enables autonomous detection and spatial analysis of anomalous behavior in crowded scenes to enhance safety in public spaces. Existing anomaly detection approaches often focus solely on event classification and provide limited information about the spatial distribution of anomalous individuals and their impact on crowd behaviour. The system, proposed in this study, integrates a customized CNN architecture for anomaly classification with the YOLOv8 algorithm, which maps crowd dynamics. This allows the system to not only detect the presence of anomalies but also simultaneously analyze the locations and behavioral statuses of individuals exhibiting abnormal behavior, thus offering a more comprehensive decision-making support mechanism in terms of spatial awareness and crowd safety. UCSD Ped2 and CUHK Avenue datasets were used in the experimental process; the data imbalance problem was addressed through random subsampling, optimizing the model's sensitivity. The developed architecture distinguishes between normal and anomalous situations by learning spatial features from $200 \times 200$ resolution video frames, while the YOLOv8 module colors individuals according to their behavior types (Normal, Panic, Calm) and reflects this on the simulation map. The obtained 90% accuracy rate and 0.89 AUC score demonstrate that the proposed method offers an effective decision support mechanism in complex scenes and crowd management.